February 27, 2025
Featured Speakers: Akylas Stratigakos, Research Associate, Imperial College London; Wangkun Xu, Research Associate, Imperial College London
Short-term energy forecasting, from a few minutes to a few days ahead, is critical for the near real-time operations of low-carbon power systems, enabling operators to better cope with variable and uncertain renewable production. Energy forecasting tools deployed in operational processes are increasingly based on machine learning methods, aggregating heterogeneous data from various sources, such as production measurements and weather forecasts. The resultant forecasts are subsequently used as input in decision-making tasks, such as market-clearing or scheduling processes. This talk presents novel ideas to improve the reliability of energy forecasts and operational decisions concerning two critical aspects: (i) dealing with missing data after model deployment, and (ii) tailoring energy forecasts to downstream decision-making tasks. In the first part, we consider the problem of missing data after a model has been deployed in production, which could result from an equipment failure or cyberattack, and we present novel forecasting methods that seamlessly handle missing data operationally by adapting to the available information. In the second part, we consider the issue of tailoring energy forecasts to the downstream decision tasks. Instead of solely focusing on statistical accuracy, we present a smart predict-and-optimize (SPO) approach that embeds knowledge about the downstream decision task during model training, leading to more economical and robust decisions.